GEO & AI Search KPIs: How to Measure AI Visibility

Five KPIs cover AI visibility end to end: whether you are cited, how prominently, how accurately, whether it sends visitors, and whether it moves pipeline. Each one answers a different question, uses a different method, and draws on a different data source. These are the KPIs for AEO and GEO alike; formulas are in each KPI section below.

Share of voice is not a sixth KPI; it is Citation Share and Answer Presence measured against named competitors, and it gets its own section below because it is how most leadership teams ask the question.

Key Takeaway: One question per KPI, one method per question. If a metric on your dashboard cannot be mapped to one of these five rows, it is probably noise.

For the narrower measurement question, see our evidence-led guide to tracking brand mentions in AI answers, including what can and cannot be inferred from a mention.

Measure AEO with the same five KPIs, putting more weight on Answer Presence and Entity Accuracy across a fixed set of buyer prompts. AEO is the answer-focused layer of GEO, so the question is whether you are included, and described correctly, when those questions are answered directly. Keep the prompt set fixed so month-to-month changes reflect visibility rather than new questions. For what AEO work involves, see the Answer Engine Optimization guide. For Google's AI features, Google documents that AI Overviews and AI Mode traffic is included in Search Console's Web search reporting (Google Search Central: AI features and your website). No KPI target or benchmark is guaranteed.

Rankings, backlinks, and impressions were built for a world where a human clicked a blue link. AI answer engines collapse the funnel: a citation appears, the answer is read, and the click often never happens. That is why a site can see flat or declining organic sessions while brand mentions, demo requests, and inbound pipeline are growing. The measurement instrument is incomplete, not necessarily the marketing.

Three specific gaps make classic SEO KPIs insufficient on their own for GEO and AI search:

• Partial referrer data. Some AI assistants pass a referrer and now appear in Google Analytics as their own channel, but many AI-influenced visits still arrive later as direct or branded-search traffic with no trace of the answer that prompted them.
• Impressions without context. Search Console now reports how often your pages appear in Google's AI Overviews and AI Mode, but an impression does not tell you whether the answer named you, recommended a competitor, or described you accurately, and it covers only Google's surfaces, not ChatGPT, Perplexity, Claude, Grok, or Copilot.
• No verbatim tracking. The signal that matters most (is the engine describing your brand, positioning, and category correctly) is not reported by any standard analytics platform.

The measurement stack in this guide fills those gaps with a parallel system: five KPIs, tracked monthly against a fixed prompt panel, read alongside the Google reports and correlated against the pipeline metrics your CEO already trusts. If you are still deciding which discipline you are measuring, the AEO vs GEO comparison explains where answer-engine and generative-engine work overlap; the KPIs below apply to both.

Key Takeaway: If your only dashboard is rankings and organic sessions, you are missing how buyers increasingly research inside AI answers. Build the parallel AI-search stack now, rather than inheriting the diagnostic problem after pipeline softens.

Every GEO and AI search KPI worth tracking falls into one of five categories. Track one metric per category monthly, and you have a defensible dashboard. Track everything and you have noise.

The measurement pyramid at a glance

The rest of this guide walks each category: method, how to read your own numbers, free tool, and the common mistakes that make the number lie to you. If you have not yet built the discovery layer these KPIs measure, start with the Generative Engine Optimization pillar and the Answer Engine Optimization pillar. For the strategic frame on why zero-click surfaces are eating clicks in the first place, see the zero-click marketing playbook. For the earned media work that creates the evidence these KPIs track, see the PR strategy guide.

Key Takeaway: Five categories, five metrics, one dashboard. Anyone selling you a 40-metric AI-visibility scorecard is selling complexity, not clarity.

Citation Share is the percentage of tracked AI responses that cite one of your URLs as a source. Responses that name your brand without citing you count toward mention rate instead; report the two separately. It is the foundation AI-search KPI because it is the most direct proxy for share of voice on the surface where buyers increasingly research.

The formula

Citation Share (source-citation rate) = (Responses Citing One of Your URLs ÷ Total Tracked Responses) × 100

A tracked response is one run of one prompt on one engine. Use the same unit for every rate in this guide: mention rate = responses naming your brand ÷ total tracked responses; source-citation rate = responses citing your URL ÷ total tracked responses; competitive share = your mentions (or citations) ÷ mentions (or citations) of you plus a fixed competitor set, on the same responses. Record the date, engine and model version, location, and whether you were signed in, and keep them constant between periods.

For example, run a panel of 20 prompts across ChatGPT, Perplexity, Gemini, Claude and Grok once each. 100 tracked responses. If 32 of them cite one of your URLs, your Citation Share is 32%. With three runs per prompt and engine, the denominator becomes 300. Add Microsoft Copilot and Google AI Mode to the panel if your buyers use them; just keep the engine list fixed once you start.

The prompt panel

Build a fixed 20-prompt panel with four classes, five prompts each:

1. Category definition: "what is [category]", "how does [category] work"
2. Vendor shortlist: "best [category] tools for [ICP]", "top [category] agencies"
3. Comparison: "X vs Y for [use case]", "alternatives to [competitor]"
4. Buyer-question: "how do I [job-to-be-done]", "what should I look for in [category]"

Run the same panel monthly. Log citation presence per engine in a spreadsheet or use the Free AI Visibility Audit, which runs a scored buyer-prompt panel across the major engines in a single pass, no signup, no credit card.

How to read your number

There is no published industry benchmark for Citation Share; results depend heavily on category, prompt wording, and which engines you track. Smart Money Media uses the following working bands as an internal planning heuristic. They are SMM methodology for deciding what to fix next, not industry standards:

• Rarely cited: Check foundations first: that AI crawlers can reach and index your key pages, and that your site states clearly who you are and what you do. Optional steps such as Organization schema, a Wikidata entry or an llms.txt file are low-cost experiments, not proven universal fixes.
• Cited on some prompts: Content and third-party citation signal are the next lever (pillar pages, tier-1 editorial).
• Cited on most shortlist prompts: You are a genuine category contender; track competitors closely.
• Cited on nearly every category prompt: Defensive strategy matters as much as offensive from this point.

For an independent reference point, our AI Citation Gap study publishes measured citation rates for funded B2B vendors on buyer-intent prompts, with the methodology and confidence intervals alongside.

Key Takeaway: Citation Share is the north-star KPI for GEO and AI search. Track it monthly with a fixed prompt panel. The month-over-month trend, compared against your own baseline, is what tells you if the work is working.

Citation Share tells you whether you appear. Answer Presence tells you how prominently you appear when you do. Not all citations are equal: a footnote link at position 8 of 10 is worth a fraction of being named in the first sentence of the answer body.

Cited vs mentioned vs quoted

Score every appearance on a three-tier scale. The weights below are Smart Money Media's scoring convention (a consistent internal method, not a universal standard), so pick weights you can defend and keep them fixed:

• Named in answer body (3 pts): engine describes your brand by name in the paragraph itself.
• Linked as top-3 source (2 pts): cited in the top three footnotes.
• Linked as lower-tier source (1 pt): cited but buried at position 4+.

Aggregate the score across the prompt panel monthly. The direction of the score matters more than the absolute number; a Citation Share holding flat while the Presence score rises means your existing citations are getting more valuable.

What moves Answer Presence

Being named in the answer body usually depends on how clearly and consistently your brand is associated with the category across sources the engines trust: editorial coverage, reference pages, and your own answer-ready content. Being a top-cited source depends more on whether your page directly and concisely answers the prompt. Track them separately: a brand that is often linked but rarely named has a positioning problem, not a content problem.

Key Takeaway: Being cited is table stakes. Being named prominently is the actual objective, and it is a separately measurable KPI.

Entity Accuracy and Sentiment measure whether an appearance is helping or hurting you. An engine that cites you but places you in the wrong category, or describes you in cautionary terms, can do more damage than not citing you at all.

Entity Accuracy: is the engine describing you correctly?

Score each engine's unprompted description of your brand on three dimensions:

1. Category correct? Does it place you in the right industry vertical?
2. Positioning correct? Does it describe your differentiator accurately?
3. Facts correct? Founding year, founder names, headquarters, and any numeric claims.

1 point per dimension, per engine. With five engines the maximum is 15 (3 dimensions × 5 engines); scale the maximum to however many engines you track. In SMM's scoring method, repeated misses on the same dimension usually point to an entity-graph problem: typically inconsistent brand naming across LinkedIn, Crunchbase, and Wikidata, or missing Organization schema.

Sentiment: are AI descriptions positive, neutral, or negative?

Rare but critical. Log any AI answer that describes your brand in negative or cautionary terms, note the sources the engine cites for that view, and track whether the same framing repeats across engines. A negative unprompted description on a buyer-intent prompt reaches prospects at exactly the moment they are forming a shortlist, which is why it deserves attention even when it appears on a single engine. Address anything demonstrably wrong at its source; the correction process is documented in the personal reputation management playbook.

Key Takeaway: Being cited prominently, accurately, and positively are three different outcomes. Score accuracy and log sentiment every month alongside Citation Share.

AI share of voice (often called share of model) is how often AI engines mention or cite your brand for a set of category prompts, compared with named competitors. It is the question leadership teams usually ask ("are we showing up more than they are?"), and it is built from the same prompt panel as Citation Share and Answer Presence.

The four measures, and how they differ

• Mention rate: tracked responses that name your brand anywhere in the answer text, whether or not they link to you, divided by total tracked responses.
• Source-citation rate (Citation Share, KPI 1): tracked responses that link to or footnote one of your URLs as a source, divided by total tracked responses. A brand can be mentioned without being cited, and cited without being mentioned.
• Answer presence: how prominently you appear when you do (KPI 2), which separates a passing footnote from a direct recommendation.
• Share of voice / share of model: your mentions or citations divided by the total mentions or citations for you and a fixed competitor set on the same prompts. See the share of model glossary entry for the definition.

A simple way to calculate it

AI Share of Voice = Your Mentions ÷ (Your Mentions + Competitor Mentions) × 100, counted on the same tracked responses (same prompts, engines, runs and date). Calculate a citation-based version the same way with citations. Report it by engine as well as in total; a strong position in one engine can hide a gap in another.

Some teams weight mentions by prominence using the Answer Presence scale above. That weighting is Smart Money Media's proprietary methodology, not a universal standard, useful for trend-tracking inside one program, but not comparable with scores from other tools or agencies.

Why prompt and query sampling matters

AI answers vary by wording, by engine, by location, and from one run to the next. A share-of-voice figure is only as good as the sample behind it:

• Use real buyer language. Seed prompts from sales calls, People Also Ask data, and query fan-out research rather than your own marketing vocabulary.
• Cover the whole journey. Include definition, shortlist, comparison, and problem-led prompts, not only branded or "best agency" head terms, which are rarely where qualified buyers start.
• Run each prompt more than once. Repeated runs per prompt and engine reduce the effect of one-off answers; report the number of runs alongside the result.
• Hold the panel and competitor set constant. Changing either mid-period breaks comparability.

Share of voice is where GEO overlaps with the broader discipline of earning visibility across every large language model; the LLM SEO guide covers the underlying work that moves it.

Key Takeaway: Report mention rate and citation rate separately, compare against a fixed competitor set, and always state the sample behind the number.

Google Search Console now includes a Generative AI performance report that shows how often your pages appear in AI Overviews and AI Mode, including which pages received those impressions. Google has rolled the report out to all websites, and the same impressions are also included in the standard Search performance report, so your overall totals already contain them (Google Search Console Help: Generative AI performance report).

What the report shows

• Impressions: how often links to your site were shown in a generative AI feature on Google Search.
• Pages: which of your URLs appeared, and which appeared most or least.
• Countries and devices: where the visibility originates.
• Dates: trends over time.

How to use it alongside the five KPIs

1. Compare AI-feature impressions with total impressions by page. Pages with a growing share of AI-feature impressions but flat clicks are the ones where the answer is satisfying the query, the place to add a clear reason to click.
2. Match cited pages to your prompt panel. If a page gains AI Overview impressions, check whether it is also cited in ChatGPT, Perplexity, and Gemini for the related prompts.
3. Remember what it does not show. Impressions do not tell you whether you were named, recommended, or described accurately, and they cover only Google's surfaces; keep Answer Presence and Entity Accuracy in your own panel.

Search Console also includes a Search generative AI control under Settings that lets site owners include or exclude their content from these features. Inclusion is the default; excluding your site removes both the impressions and the traffic from these features, so treat it as a deliberate business decision rather than a measurement setting.

Key Takeaway: Google now reports AI Overview and AI Mode impressions by page. Use it as your Google-side visibility signal, and keep your prompt panel for everything Google does not report.

Google Analytics now groups recognized AI assistant referrals into a default "AI Assistant" channel, assigning them the medium "ai-assistant", so the measurable part of AI traffic no longer has to be dug out of Referral by hand. According to Google, the channel covers visits from sources such as ChatGPT, Gemini, DeepSeek, Copilot, and Grok, and excludes Google's AI Overviews and AI Mode, which are reported as Google organic search (Google Analytics Help: Default channel group).

That makes referral data a floor, not a total. Many people who see your brand in an AI answer never click the citation; they search your name later, type your URL, or come back through another channel. A defensible attribution read combines the channel data with the downstream signals below.

Step 1: Review the AI Assistant channel

In Google Analytics, open Reports → Acquisition → Traffic acquisition and set the primary dimension to Session default channel group. Look at AI Assistant sessions, engaged sessions, and key events month over month.

Step 2: Break it down by source/medium

Switch the dimension to Session source / medium to see which assistants send traffic: for example chatgpt.com, perplexity.ai, gemini.google.com, copilot.microsoft.com, and claude.ai. If you previously built a custom channel group for AI referrals, compare it with the default channel; Google has not published a complete list of recognized sources, so a custom group can still catch assistants the default misses.

Step 3: Check landing pages and engagement

Add Landing page as a secondary dimension. The pages receiving AI assistant visits are usually the pages being cited; compare them with the URLs your prompt panel and Search Console's generative AI report show as cited. Check engagement rate and engagement time: visitors arriving from an AI answer often land deep in the site with specific intent.

Step 4: Tie it to conversions and key events

Report key events (audit starts, contact forms, booked calls) by channel. Compare the AI Assistant channel's conversion rate with organic search on your own data rather than assuming it will be higher or lower; samples are often small, so read quarterly totals, not single months.

Step 5: Track branded-search lift and assisted demand

AI visibility often shows up as more people searching for your brand by name. Track branded query impressions and clicks in Search Console monthly and compare them with your Citation Share trend, and watch direct traffic to the pages that are being cited. Our guide to tracking branded search lift covers the method in detail.

If you want a single "attributed AI sessions" number, a simple SMM working formula is:

Attributed AI Sessions = AI Assistant Channel Sessions + (Direct-Traffic Delta to Cited URLs) + (Branded Search Delta × Your Chosen Attribution Share)

The attribution share is an assumption you choose and document, not a published coefficient. Keep it conservative, hold it constant, and treat the result as a trend indicator rather than a precise count.

Measure this against your existing tools: a pipeline-value model layers business value on top of session counts, and a citation panel tells you which of your pages are actually being cited in the first place. Explore the full measurement toolkit for pipeline-value reporting.

Key Takeaway: Start with the AI Assistant channel, then read landing pages, key events, and branded-search lift alongside it. Referral data shows the clicks; it will never show all of the influence.

Every GEO and AI search program eventually gets asked one question: what did this cost us versus what did it earn us. Two formulas and one sales-process habit answer it, and all three belong in your quarterly review.

Cost per Citation

Cost per Citation = Total Program Cost ÷ Net New Citations Earned

Program cost includes agency retainers, in-house comms salaries allocated to AI-visibility work, editorial content production, and any tool subscriptions. Net new citations = the count of unique cited URLs across your tracked engines this quarter minus last quarter.

There is no reliable public benchmark for cost per citation; costs vary widely by category, competition, and starting authority. Use your own first two quarters as the baseline and judge the trend. A cost per citation that keeps rising usually means the foundational entity work has not been done, and you are paying content-production prices for signals that will not compound.

Earned Media ROI (AI-adjusted)

The classic earned-media ROI formula (media value ÷ program cost) breaks in AI search because there is no CPM equivalent for an AI citation. The AI-adjusted version:

AI-Adjusted ROI = (Attributed Sessions × Conv. Rate × Avg Deal Size) ÷ Program Cost

Run this quarterly, not monthly; the sample size at 30 days is too small to be reliable. For the financial side, the Earned Media ROI Calculator computes it from inputs most operators already have; treat media-cost-equivalent estimates and financial ROI as separate from GEO and AI visibility metrics, never as a substitute for them. For the zero-click side of the same calculation, see measuring zero-click marketing ROI.

The lagging indicator: influenced pipeline

Sales teams that ask "how did you hear about us" during discovery can tag deals with an AI-influenced flag, and add "ChatGPT", "Perplexity", "Google AI answer" and similar options to the self-reported attribution field on your forms. After at least six months of tagging, compare win rate, deal size, and sales-cycle length on AI-influenced deals against your own baseline. Whatever the comparison shows, it is your evidence, not an industry average.

Set expectations accordingly: meaningful movement in AI visibility and its pipeline effect commonly takes 6–12 months or longer, depending on starting position, execution, competition, indexing, and platform changes. No KPI framework can guarantee citations, traffic, or revenue.

Key Takeaway: ROI is the KPI that saves the budget. Cost per Citation for the operating team, AI-Adjusted ROI for finance, influenced-pipeline comparison for the CEO. Report all three quarterly.

You do not need an enterprise AI-search platform to measure any of the KPIs in this guide. The following free tools cover the full stack, together with the free Google reports above; build the measurement discipline first, buy enterprise tooling only when volume demands it.

Every tool above is free, requires no signup for basic use, and outputs the exact inputs the KPI formulas in this guide require. Search Console's Generative AI performance report and Google Analytics' AI Assistant channel are free as well.

Key Takeaway: The free tool stack in this table is the fastest way to stand up a defensible AI-search measurement program. Discipline beats tooling. Start monthly, hold the methodology constant, and let the trend lines do the talking.

KPIs that get measured but never reviewed are worse than KPIs that do not exist; they create false confidence. The cadence below is what separates programs that compound from programs that plateau.

The first business day of every month

• Run the fixed 20-prompt panel across all tracked engines. Log Citation Share, Presence score, Entity Accuracy score, sentiment notes, and AI share of voice against your competitor set.
• Update the master spreadsheet or dashboard. Never change the prompt panel mid-quarter; you lose comparability.
• Address any new negative or inaccurate unprompted descriptions at their source.

The 15th of every month

• Pull the AI Assistant channel, landing pages, and key events from Google Analytics; pull the Generative AI performance report and branded-search data from Search Console. Compute Attributed AI Sessions.
• Compare against the prior month's Citation Share trend. A rising Citation Share with flat sessions means the cited pages may be the wrong ones; fix content or internal linking.

Quarterly, in the board deck

• Cost per Citation and AI-Adjusted ROI, with quarter-over-quarter deltas.
• Influenced-pipeline comparison versus baseline.
• One qualitative slide: the three most valuable citations earned this quarter, with the buyer journey they enabled.

The one anti-pattern that kills every AI-search program

Measuring weekly. AI-search signals often move on a longer lag from the work; weekly reporting produces noise that gets misread as trend, and leads to strategy changes that reset the compound curve. Monthly is the floor, with six-month-minimum implementation horizons, quarterly ROI conversations, and an expectation that meaningful movement commonly takes 6–12 months or longer. If you would rather have this measurement run for you, our GEO agency team builds and reports this dashboard as part of the engagement.

Key Takeaway: Fixed prompt panel, monthly cadence, quarterly ROI review. Hold the methodology constant and the numbers become decisions instead of debate.